quizz_machine.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 import math, os, tqdm, warnings
   9
  10 import torch, torchvision
  11
  12 from torch import nn
  13 from torch.nn import functional as F
  14
  15 import mygpt
  16 from mygpt import BracketedSequence
  17
  18 ######################################################################
  19
  20 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
  21 # 1s where tokens should be generated. The others are kept
  22 # unchanged.
  23
  24
  25 def one_batch_masked_inplace_autoregression(
  26     model,
  27     input,
  28     ar_mask,
  29     seq_logproba,
  30     temperature=1.0,
  31     deterministic_synthesis=False,
  32     forbidden_tokens=None,
  33     forced_biases=None,
  34 ):
  35     to_generate = (ar_mask.sum(0) > 0).nonzero()
  36
  37     if to_generate.min() > 0:
  38         model(
  39             BracketedSequence(input, 0, to_generate.min())
  40         )  # Needed to initialize the model's cache
  41     for s in range(to_generate.min(), to_generate.max() + 1):
  42         output = model(BracketedSequence(input, s, 1)).x
  43
  44         logits = output[:, s]
  45
  46         logits = (logits / temperature).log_softmax(dim=-1)
  47
  48         if forbidden_tokens is not None:
  49             logits = logits.masked_fill(forbidden_tokens, float("-inf"))
  50
  51         if forced_biases is not None:
  52             logits = logits + forced_biases[None, :]
  53
  54         if deterministic_synthesis:
  55             t_next = logits.argmax(-1)
  56         else:
  57             dist = torch.distributions.categorical.Categorical(logits=logits)
  58             t_next = dist.sample()
  59
  60         all_n = torch.arange(t_next.size(0))
  61         seq_logproba += logits[all_n, t_next].sum(dim=-1)
  62
  63         input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
  64
  65
  66 def masked_inplace_autoregression(
  67     model,
  68     batch_size,
  69     input,
  70     ar_mask,
  71     seq_logproba,
  72     temperature,
  73     deterministic_synthesis,
  74     forbidden_tokens=None,
  75     logit_biases=None,
  76     progress_bar_desc=None,
  77     device=torch.device("cpu"),
  78 ):
  79     assert input.size() == ar_mask.size()
  80
  81     batches = zip(
  82         input.split(batch_size),
  83         ar_mask.split(batch_size),
  84         seq_logproba.split(batch_size),
  85     )
  86
  87     if progress_bar_desc is not None:
  88         batches = tqdm.tqdm(
  89             batches,
  90             dynamic_ncols=True,
  91             desc=progress_bar_desc,
  92             total=(input.size(0) + batch_size - 1) // batch_size,
  93         )
  94
  95     with torch.autograd.no_grad():
  96         t = model.training
  97         model.eval()
  98
  99         for input, ar_mask, seq_logproba in batches:
 100             one_batch_masked_inplace_autoregression(
 101                 model=model,
 102                 input=input,
 103                 ar_mask=ar_mask,
 104                 seq_logproba=seq_logproba,
 105                 temperature=temperature,
 106                 deterministic_synthesis=deterministic_synthesis,
 107                 forbidden_tokens=forbidden_tokens,
 108                 forced_biases=logit_biases,
 109             )
 110
 111         model.train(t)
 112
 113
 114 ######################################################################
 115
 116
 117 class QuizzMachine:
 118     def make_ar_mask(self, input):
 119         b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
 120         return b.long()[None, :].expand_as(input)
 121
 122     def __init__(
 123         self,
 124         problem,
 125         nb_train_samples,
 126         nb_test_samples,
 127         batch_size,
 128         result_dir,
 129         logger,
 130         device=torch.device("cpu"),
 131     ):
 132         super().__init__()
 133
 134         self.problem = problem
 135         self.batch_size = batch_size
 136         self.device = device
 137         self.logger = logger
 138
 139         self.train_w_quizzes = self.problem.generate_token_sequences(
 140             nb_train_samples
 141         ).to(device)
 142         self.test_w_quizzes = self.problem.generate_token_sequences(nb_test_samples).to(
 143             device
 144         )
 145
 146         self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
 147
 148         self.train_c_quizzes = []
 149         self.test_c_quizzes = []
 150
 151         if result_dir is not None:
 152             self.problem.save_quizzes(
 153                 self.train_w_quizzes[:72], result_dir, "culture_w_quizzes"
 154             )
 155
 156     def batches(self, split="train", desc=None):
 157         assert split in {"train", "test"}
 158         if split == "train":
 159             w_quizzes = self.train_w_quizzes
 160             c_quizzes = self.train_c_quizzes
 161         else:
 162             w_quizzes = self.test_w_quizzes
 163             c_quizzes = self.test_c_quizzes
 164
 165         if len(c_quizzes) > 0:
 166             c_quizzes = torch.cat(c_quizzes, dim=0)
 167             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 168                 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
 169                 c_quizzes = c_quizzes[i]
 170
 171             i = torch.randperm(w_quizzes.size(0))[
 172                 : w_quizzes.size(0) - c_quizzes.size(0)
 173             ]
 174             w_quizzes = w_quizzes[i]
 175
 176             self.nb_batch_w_quizzes = w_quizzes.size(0)
 177             self.nb_batch_c_quizzes = c_quizzes.size(0)
 178
 179             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 180         else:
 181             input = w_quizzes
 182             self.nb_batch_w_quizzes = w_quizzes.size(0)
 183             self.nb_batch_c_quizzes = 0
 184
 185         # Shuffle
 186         input = input[torch.randperm(input.size(0))]
 187
 188         if desc is None:
 189             desc = f"epoch-{split}"
 190         for batch in tqdm.tqdm(
 191             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 192         ):
 193             yield batch
 194
 195     def vocabulary_size(self):
 196         return self.nb_codes
 197
 198     def produce_results(
 199         self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
 200     ):
 201         def compute_accuracy(input):
 202             input = input[:nmax]
 203             ar_mask = self.make_ar_mask(input)
 204             result = input.clone() * (1 - ar_mask)
 205             seq_logproba = torch.empty(input.size(0), device=self.device)
 206
 207             masked_inplace_autoregression(
 208                 model=model,
 209                 batch_size=self.batch_size,
 210                 input=result,
 211                 ar_mask=ar_mask,
 212                 seq_logproba=seq_logproba,
 213                 temperature=1.0,
 214                 deterministic_synthesis=deterministic_synthesis,
 215                 progress_bar_desc=None,
 216                 device=self.device,
 217             )
 218
 219             nb_total, nb_correct = (
 220                 input.size(0),
 221                 (input == result).long().min(dim=1).values.sum(),
 222             )
 223
 224             return nb_total, nb_correct
 225
 226         train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
 227
 228         self.logger(
 229             f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 230         )
 231
 232         test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
 233
 234         self.logger(
 235             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 236         )
 237
 238         main_test_accuracy = test_nb_correct / test_nb_total
 239         self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 240
 241         ##############################
 242
 243         input = self.test_w_quizzes[:96]
 244         ar_mask = self.make_ar_mask(input)
 245         result = input.clone() * (1 - ar_mask)
 246         seq_logproba = torch.empty(input.size(0), device=self.device)
 247
 248         masked_inplace_autoregression(
 249             model=model,
 250             batch_size=self.batch_size,
 251             input=result,
 252             ar_mask=ar_mask,
 253             seq_logproba=seq_logproba,
 254             temperature=1.0,
 255             deterministic_synthesis=deterministic_synthesis,
 256             progress_bar_desc=None,
 257             device=self.device,
 258         )
 259
 260         self.problem.save_quizzes(
 261             result[:72], result_dir, f"culture_prediction_{n_epoch:04d}_{model.id:02d}"
 262         )
 263
 264         return main_test_accuracy
 265
 266     def renew_w_quizzes(self, nb, for_train=True):
 267         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 268         nb = min(nb, input.size(0))
 269         input[:-nb] = input[nb:].clone()
 270         input[-nb:] = self.problem.generate_token_sequences(nb).to(self.device)
 271
 272     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 273         if for_train:
 274             self.train_c_quizzes.append(new_c_quizzes)
 275         else:
 276             self.test_c_quizzes.append(new_c_quizzes)
 277
 278     def reverse_time(self, c_quizzes):
 279         token_forward, token_backward = self.problem.direction_tokens()
 280
 281         l = (c_quizzes.size(1) - 1) // 2
 282         direction = c_quizzes[:, l : l + 1]
 283         direction = self.problem.token_forward * (
 284             direction == self.problem.token_backward
 285         ) + self.problem.token_backward * (direction == self.problem.token_forward)
 286
 287         return torch.cat([c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1)
 288
 289     def compute_correctness(
 290         self, c_quizzes, models_for_validation, both_direction=True
 291     ):
 292         reversed_c_quizzes = self.reverse_time(c_quizzes)
 293
 294         ar_mask = self.make_ar_mask(c_quizzes)
 295         seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
 296
 297         # Check how many of models can solve the quizzes in both directions
 298
 299         nb_correct = 0
 300
 301         for model in models_for_validation:
 302             result = c_quizzes.clone()
 303
 304             masked_inplace_autoregression(
 305                 model=model,
 306                 batch_size=self.batch_size,
 307                 input=result,
 308                 ar_mask=ar_mask,
 309                 seq_logproba=seq_logproba,
 310                 temperature=1.0,
 311                 deterministic_synthesis=True,
 312                 # progress_bar_desc="solving c_quizzes",
 313                 device=self.device,
 314             )
 315
 316             correct = (c_quizzes == result).long().min(dim=-1).values
 317
 318             if both_direction:
 319                 reversed_result = reversed_c_quizzes.clone()
 320
 321                 masked_inplace_autoregression(
 322                     model=model,
 323                     batch_size=self.batch_size,
 324                     input=reversed_result,
 325                     ar_mask=ar_mask,
 326                     seq_logproba=seq_logproba,
 327                     temperature=1.0,
 328                     deterministic_synthesis=True,
 329                     # progress_bar_desc="solving reversed c_quizzes",
 330                     device=self.device,
 331                 )
 332
 333                 reversed_correct = (
 334                     (reversed_c_quizzes == reversed_result).long().min(dim=-1).values
 335                 )
 336
 337                 correct *= reversed_correct
 338
 339             # endif
 340
 341             nb_correct += correct
 342
 343         return nb_correct
 344
 345     ###############################################################
 346
 347     def generate_quizzes(self, nb, model_for_generation, reverse_cleanup=False):
 348         c_quizzes = torch.empty(
 349             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 350         )
 351
 352         ar_mask_prompt = torch.zeros(c_quizzes.size(), device=self.device)
 353         ar_mask_prompt[:, : ar_mask_prompt.size(1) // 2 + 1] = 1
 354         ar_mask_solve = 1 - ar_mask_prompt
 355         seq_logproba = torch.empty(ar_mask_prompt.size(0), device=self.device)
 356
 357         if reverse_cleanup:
 358             warnings.warn("very high temperature with reversed cleanup", RuntimeWarning)
 359             temperature = 10.0
 360         else:
 361             temperature = 1.0
 362
 363         # warnings.warn("noise injection", RuntimeWarning)
 364         # noise_std = torch.rand(1).item()
 365         # self.logger(f"{noise_std=}")
 366
 367         # mygpt.set_noise_injection(model_for_generation, noise_std)
 368
 369         masked_inplace_autoregression(
 370             model=model_for_generation,
 371             batch_size=self.batch_size,
 372             input=c_quizzes,
 373             ar_mask=ar_mask_prompt,
 374             seq_logproba=seq_logproba,
 375             temperature=temperature,
 376             deterministic_synthesis=False,
 377             device=self.device,
 378         )
 379
 380         # mygpt.set_noise_injection(model_for_generation, 0.0)
 381
 382         ave_seq_logproba = seq_logproba.mean()
 383
 384         masked_inplace_autoregression(
 385             model=model_for_generation,
 386             batch_size=self.batch_size,
 387             input=c_quizzes,
 388             ar_mask=ar_mask_solve,
 389             seq_logproba=seq_logproba,
 390             temperature=temperature,
 391             deterministic_synthesis=True,
 392             device=self.device,
 393         )
 394
 395         if reverse_cleanup:
 396             c_quizzes = self.reverse_time(c_quizzes)
 397             masked_inplace_autoregression(
 398                 model=model_for_generation,
 399                 batch_size=self.batch_size,
 400                 input=c_quizzes,
 401                 ar_mask=ar_mask_solve,
 402                 seq_logproba=seq_logproba,
 403                 temperature=temperature,
 404                 deterministic_synthesis=True,
 405                 device=self.device,
 406             )
 407
 408             c_quizzes = self.reverse_time(c_quizzes)
 409             masked_inplace_autoregression(
 410                 model=model_for_generation,
 411                 batch_size=self.batch_size,
 412                 input=c_quizzes,
 413                 ar_mask=ar_mask_solve,
 414                 seq_logproba=seq_logproba,
 415                 temperature=temperature,
 416                 deterministic_synthesis=True,
 417                 device=self.device,
 418             )
 419
 420         return c_quizzes, seq_logproba.mean()